Cohere has open-sourced Transcribe Arabic, a 2B-parameter speech-to-text model that now leads the Hugging Face Arabic ASR leaderboard at 25.87 WER (Word Error Rate)—approximately 11 points better than Whisper Large V3.
Performance Breakthrough
Transcribe Arabic achieves state-of-the-art results on Arabic speech recognition: - 25.87 WER — leaderboard #1 - 2B parameters — Apache 2.0 licensed - 96% human preference — in head-to-head tests against Whisper
The model handles dialectal variety, code-switching between Arabic and English, and supports mlx-audio for on-device inference from day zero.
Addressing Arabic Speech Recognition Challenges
Arabic presents unique challenges for ASR systems due to: - Multiple dialects (Egyptian, Gulf, Levantine, Maghrebi, Modern Standard Arabic) - Code-switching between Arabic and English - Diacritical marks and morphological complexity
Transcribe Arabic addresses these challenges by training on diverse Arabic speech data, enabling robust performance across dialects and conversational contexts.
Open Source Strategy
The Apache 2.0 license enables commercial use, making it accessible for enterprises building Arabic-language applications. This release positions Cohere as a leader in multilingual speech recognition while contributing to the open-source ecosystem.
The model’s strong performance (11 points better than Whisper) demonstrates that specialized models can significantly outperform general-purpose solutions in specific languages—a trend likely to continue as the industry moves toward specialized rather than generalist AI systems.